Source-linked AI summary

AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels

Javier Tirado-Garín, Alan Savio Paul, Shuai Chen, Axel Barroso-Laguna, Tommaso Cavallari, Daniyar Turmukhambetov, Victor Adrian Prisacariu, Eric Brachmann

arXiv:2609.02798v1cs.CV

TL;DR

Neural map matchers are trained with geo-referenced pose labels that often contain inaccuracies. AutoCompass learns from weak supervision using GPS positions and relative poses, and consistently outperforms strongly supervised methods across driving and egocentric benchmarks.

  • Problem

    Large-scale geo-referenced image datasets contain noisy absolute position and heading labels, limiting reliable supervision for neural map matchers.

  • Method

    AutoCompass uses tolerance regions around GPS positions, omits heading labels, and incorporates relative-pose supervision to train map matchers from weak labels.

  • Results

    AutoCompass consistently outperforms strongly supervised methods across driving and egocentric benchmarks, with relative-pose supervision achieving a new state of the art.

  • Takeaways & Limitations

    Neural map matchers can be trained accurately using raw GPS positions or relative poses instead of accurate geo-referenced 3-DoF labels.

  • Takeaways & Limitations

    Relative translation supervision requires ground-truth translations expressed in the map’s geo-referenced coordinate system.

Abstract

from arXiv · show

Neural map matchers estimate an image's 3-DoF pose relative to a 2D map. These models are trained on large-scale datasets of geo-referenced images, whose position and heading labels often contain noise that affects the trained models. To address this, we present AutoCompass, a supervision approach for training neural map matchers from inaccurate absolute pose labels. First, we show that heading labels are unnecessary: trained from raw GPS labels, models learn to predict accurate headings, automatically. Second, defining a tolerance region around raw GPS improves positional accuracy. Third, if available, our supervision uses relative poses between training images, obtained via SLAM or SfM, which provide a more accurate training signal. Across driving and egocentric benchmarks, AutoCompass consistently outperforms counterparts trained with the usual strong reliance on absolute pose labels.

1 Introduction

AutoCompass addresses noisy absolute pose supervision in neural map matching by eliminating heading labels, tolerating GPS position errors, and exploiting relative poses. Across multiple datasets, it outperforms strong-supervision baselines and achieves state-of-the-art visual localization on public 2D maps.

  • Motivation: The method targets noise in large-scale geo-referenced training data, where automated absolute position and heading annotations are susceptible to inaccuracies.Prior approaches commonly combine local SfM reconstructions with raw GPS measurements.
  • Method: AutoCompass removes the need for geo-referenced heading labels and assumes the true image position lies within a tolerance region of up to 20 m around GPS coordinates.The approach relies on architectural inductive bias for accurate heading prediction and increases robustness to ground-truth inaccuracies.
  • Method: AutoCompass provides absolute and practical relative pose losses, including variants for non-metric or approximately geo-referenced poses, to improve existing single-image supervised approaches.Relative poses can be estimated reliably using SfM or SLAM.
  • Results: Raw GPS-trained AutoCompass models outperform strong-supervision baselines across multiple datasets, while relative pose supervision achieves state-of-the-art visual localization on public 2D maps.The method remains applicable to existing architectures without modification.

2 Related work

Cross-view visual localization estimates 3-DoF poses from large-scale georeferenced 2D maps, avoiding dense ground-level mapping. Recent weakly supervised methods reduce dependence on accurate pose labels, while AutoCompass uses noisy GPS or relative poses with perspective images and outperforms strongly supervised approaches.

  • Map-based localization: Neural-map matchers use lightweight, publicly available maps and neural matching, reducing the need to build and store dense ground-level 3D representations.OrienterNet uses approximately 200KB for a 128x128m area and cross-correlates learned BEV image and map features.
  • Cross-view visual localization: Cross-view methods localize images by estimating 3-DoF poses against 2D maps from aerial, satellite, or planimetric imagery, enabling wide-area localization without dense image-based maps.These maps are typically georeferenced and available at large scale.
  • Weak supervision: Weakly supervised approaches address inaccurate pose labels through emergent pose estimation, panorama-based teacher-student learning, weak position supervision, orientation priors, or label-quality adaptation.These methods relax supervision under different assumptions, including panoramas, strong orientation priors, or heterogeneous ground-truth quality.
  • AutoCompass: AutoCompass requires only perspective images plus noisy GPS measurements or relative poses for supervision and outperforms strongly supervised approaches.This contrasts with methods that assume accurate 3-DoF pose labels and addresses the effects of noisy geo-referencing.

3 Method

AutoCompass estimates a discrete 3-DoF pose distribution from a calibrated image and rasterized local map, while replacing exact absolute-pose supervision with heading marginalization, GPS-region supervision, and relative-pose losses. These mechanisms exploit geometric matching to tolerate inaccurate labels and use relative information from multi-view poses.

  • Pose estimation: AutoCompass estimates a categorical distribution over discrete 3-DoF poses by cross-correlating learned map features with rotated BEV image features.The output is an S×S×N tensor normalized with softmax over candidate positions and headings.
  • Automatic heading angle: Heading supervision is removed by marginalizing the predicted probability volume over all headings at the labeled position.Feature matching supplies a geometric inductive bias that enables heading prediction from single perspective images, with accuracy on-par with or better than heading-supervised methods.
  • Relative poses: Relative rotation and translation supervision use relative information, making them invariant to shared offsets in absolute-pose labels.Relative translation still requires the ground-truth relative translation to be expressed in the same geo-referenced coordinate system as the map tiles.
  • Learning from GPS: GPS supervision marginalizes probability mass within a ±r meter position neighborhood while also marginalizing over orientation, reducing sensitivity to biased or adverse-condition GPS errors.The method encourages probability anywhere within the GPS neighborhood, and similar performance is obtained for r ∈[5, 20] m.
  • Relative poses: Relative poses from SLAM or SfM provide training supervision without explicit geo-referencing, provided a coarse 2D location is available to select a covering map tile.Relative rotation supervision is formed by circular cross-correlation of the two images’ marginalized heading distributions and optimized at the ground-truth relative rotation.

4 Experiments

Across driving and egocentric benchmarks, AutoCompass trained with raw GPS already outperforms strongly supervised baselines, while relative-pose supervision achieves the strongest results. Its gains extend to DINOv2 backbones, unseen urban scenes, and sequential fusion.

  • Egocentric sequences: On LaMAria and ODN, AutoCompass generalizes to unseen urban scenes, outperforms alternatives trained on more data, and improves LaMAria strict-threshold performance over GPS.Using DINOv2 consistently improves over ResNet-101.
  • Overall results: Across KITTI and egocentric benchmarks, AutoCompass outperforms strongly supervised and alternative methods, with relative-pose supervision performing best and DINOv2 further improving results.The egocentric evaluation reports the relative-pose, DINOv2 variant as best across both datasets.
  • Sequential fusion: Under 10 fused frames, AutoCompass surpasses GPS recall@5m on LaMAria, whereas retrained OrienterNet requires over 20 frames.The result demonstrates that more accurate single-view estimates benefit sequential fusion.

5 Conclusion

AutoCompass enables 3-DoF visual localization on 2D maps using weak supervision from 2D GPS labels or relative poses, remaining robust to geo-referencing errors and outperforming strongly supervised methods.

  • 5 Conclusion: AutoCompass consistently outperforms strongly supervised methods while handling GPS noise and absolute pose offsets.The approach uses weak supervision rather than accurate geo-referenced 3-DoF poses.
  • 5 Conclusion: AutoCompass performs 3-DoF visual localization on 2D maps using only 2D GPS labels or relative poses.These supervision signals are easier to obtain at scale than accurate geo-referenced 3-DoF poses.

Supplementary Material

The supplementary material provides additional training and evaluation details, along with quantitative and qualitative results.

  • Supplementary Material: Sections A–D provide additional training, evaluation, quantitative, and qualitative results.Sections A and B cover training and evaluation details, while Sections C and D present additional quantitative and qualitative results.

A Additional training details

The training setup uses perturbed raw-GPS-centered map tiles, SfM-linked image pairs for relative-pose supervision, and backbone-specific image preprocessing. It follows OrienterNet’s architecture while fine-tuning pretrained backbones with separate decoder training.

  • Tile sampling: Training tiles are centered on raw GPS coordinates and randomly perturbed within ±48 m, while refined SfM coordinates are available in MGL.MGL provides raw GPS and SfM-refined coordinates; raw GPS centers the 128 × 128 m OSM tiles.
  • Relative poses: Relative-pose pairs come from images jointly optimized in the same SfM cluster, with independent augmentation corrections applied before computing relative distributions.Clusters are identified through the Mapillary merge_cc attribute, and predicted probability volumes are untransformed before relative-pose computation.
  • Images: Training images receive color jitter and are resized to 512×512 for ResNet or 518 × 518 for DINOv2 backbones.The MGL images are mostly square 90° FoV perspective crops with an average focal length of approximately 256 pixels.
  • Architecture and hyperparameters: The setup uses 64 discrete rotations, a 64.5 × 32 m BEV at 0.5 m/pixel, and separate learning rates for the pretrained backbone and DPT decoder.The DINOv2 backbone is fine-tuned at 10^-7, while the decoder is trained from scratch at 10^-5.
  • Relative poses: Relative-pose supervision derives distance, chunked distance, shift, and rotation probability distributions from paired absolute translation predictions.Figure 6 illustrates the distributions used for relative-pose supervision and notes that the raw distance distribution is non-smooth.

B Additional evaluation details

The evaluation uses standardized image preprocessing and dataset-specific protocols to address aspect-ratio mismatch, biased or locally inaccurate ground truth, and coordinate-system differences. Oxford Day-and-Night evaluation robustly aligns each model’s predictions with ground truth before computing errors.

  • General: Across evaluation datasets, images are resized to a 256-pixel focal length, with LaMAria and Oxford Day-and-Night additionally cropped to 512×512 or 518×518 pixels for DINOv2.OrienterNet, OSMLoc, and AutoCompass use 256 discrete rotations during evaluation.
  • KITTI: KITTI images are center-padded to 518×518 pixels for DINOv2-based models after focal-length resizing, while ResNet models only undergo focal-length resizing.The padding addresses KITTI’s approximately 3.3:1 aspect ratio versus the 1:1 training distribution; ResNet models benefit less from padding.
  • LaMAria: LaMAria evaluation uses eleven training sequences with pseudo-dense ground truth, excluding clips recorded while surveying ground control points.Sequences require available pseudo-dense ground truth and more than one observed control point for geo-referencing.
  • LaMAria: LaMAria headings are corrected by approximately 0.8° to convert grid north in LV95 ground truth to geodetic north in the evaluation coordinate system.The offset uses meridian convergence at the center of the topocentric system.
  • Oxford Day-and-Night: Oxford Day-and-Night evaluation robustly aligns ground truth to each model’s estimates per sequence before computing errors, addressing meter-level geo-referencing errors and single-view outliers.The alignment estimates a 3-DoF position transformation and heading offset using truncated losses, with 2 m and 3° outlier thresholds.

C Additional quantitative results

Additional evaluations show that AutoCompass’s weak chunk-marginalization supervision consistently benefits from informative relative-pose labels, while remaining robust across protocols and parameter choices. GPS neighborhood marginalization improves performance over direct GPS supervision, whereas relative-distance binning is comparatively insensitive.

  • Supervision impact: Across KITTI, LaMAria, and ODN, performance improves with more informative supervision: weak chunk marginalization outperforms strong supervision, and relative-pose labels perform best.The conclusions remain consistent across average and sequential results, although ODN shows similar performance across checkpoints.
  • KITTI evaluation with a ±10° orientation prior: Under a ±10° orientation prior, AutoCompass still improves over strongly supervised baselines, with relative-pose supervision performing best among OSM-based methods.The protocol restricts orientation search to ±10° and position search to a 40×40 m area centered on the OSM tile.
  • Varying the chunk size r for GPS-only supervision: GPS marginalization with chunk sizes r ∈[5, 20] m consistently outperforms strong supervision, while r = 0 and r = 30 m perform worse.Direct GPS supervision lacks local-neighborhood marginalization, whereas larger chunks weaken the training signal.
  • Varying the norm binning size r_∆t: Relative-distance supervision achieves similar performance across different norm binning sizes r_∆t, indicating that this parameter is not sensitive.The comparison uses the additional experiments summarized in Table 7.
  • GPS + relative pose supervision: Combining GPS with relative-pose supervision improves over GPS alone, but remains less accurate than using relative-pose supervision by itself.Increasing the supervised GPS neighborhood size improves the combined setup, unlike GPS-only supervision.

D Additional qualitative results

Qualitative comparisons across LaMAria, Oxford Day-and-Night, and KITTI show that AutoCompass learns sharper, more selective map features than strongly supervised baselines. This behavior remains consistent across the proposed weak-supervision types and becomes sharper with relative poses.

  • Comparison across methods: Across LaMAria, Oxford Day-and-Night, and KITTI, qualitative visualizations compare AutoCompass with OrienterNet and OSMLoc using relative-pose supervision.The visualizations use AutoCompass trained with ΔΘ + ΔXY supervision from Table 3.
  • Comparison across methods: AutoCompass learns sharper map features that distinguish narrow roads and alleys, focus on distinctive keypoints, and ignore uninformative regions such as large bodies of water.Feature norms track cross-correlation magnitudes, highlighting building corners and road intersections visible in the input images.
  • Comparison across types of weak supervision: The sharper-feature behavior is consistent across the different weak-supervision types and becomes even sharper when relative poses are used.The qualitative comparison suggests that relative poses and GPS-neighborhood supervision are less noisy than absolute pose labels.
Loading 2609.02798v1…